Center for Molecular Imaging and Nuclear Medicine, School of Radiological and Interdisciplinary Sciences (RAD-X), Soochow University, Collaborative Innovation Center of Radiation Medicine of Jiangsu Higher Education Institutions, Suzhou, Jiangsu, 215123, China.
School of Mechanical and Electronic Engineering, Soochow University, Suzhou, Jiangsu, 215021, China.
Med Phys. 2017 Jul;44(7):3752-3760. doi: 10.1002/mp.12350. Epub 2017 Jun 16.
Ultrasound (US) imaging has been widely used in breast tumor diagnosis and treatment intervention. Automatic delineation of the tumor is a crucial first step, especially for the computer-aided diagnosis (CAD) and US-guided breast procedure. However, the intrinsic properties of US images such as low contrast and blurry boundaries pose challenges to the automatic segmentation of the breast tumor. Therefore, the purpose of this study is to propose a segmentation algorithm that can contour the breast tumor in US images.
To utilize the neighbor information of each pixel, a Hausdorff distance based fuzzy c-means (FCM) method was adopted. The size of the neighbor region was adaptively updated by comparing the mutual information between them. The objective function of the clustering process was updated by a combination of Euclid distance and the adaptively calculated Hausdorff distance. Segmentation results were evaluated by comparing with three experts' manual segmentations. The results were also compared with a kernel-induced distance based FCM with spatial constraints, the method without adaptive region selection, and conventional FCM.
Results from segmenting 30 patient images showed the adaptive method had a value of sensitivity, specificity, Jaccard similarity, and Dice coefficient of 93.60 ± 5.33%, 97.83 ± 2.17%, 86.38 ± 5.80%, and 92.58 ± 3.68%, respectively. The region-based metrics of average symmetric surface distance (ASSD), root mean square symmetric distance (RMSD), and maximum symmetric surface distance (MSSD) were 0.03 ± 0.04 mm, 0.04 ± 0.03 mm, and 1.18 ± 1.01 mm, respectively. All the metrics except sensitivity were better than that of the non-adaptive algorithm and the conventional FCM. Only three region-based metrics were better than that of the kernel-induced distance based FCM with spatial constraints.
Inclusion of the pixel neighbor information adaptively in segmenting US images improved the segmentation performance. The results demonstrate the potential application of the method in breast tumor CAD and other US-guided procedures.
超声（US）成像已广泛应用于乳腺肿瘤的诊断和治疗干预。肿瘤的自动描绘是一个关键的第一步，特别是对于计算机辅助诊断（CAD）和 US 引导的乳腺程序。然而，US 图像的固有特性，如低对比度和模糊边界，对乳腺肿瘤的自动分割提出了挑战。因此，本研究的目的是提出一种能够描绘 US 图像中乳腺肿瘤的分割算法。
为了利用每个像素的邻域信息，采用了基于 Hausdorff 距离的模糊 C 均值（FCM）方法。通过比较它们之间的互信息，自适应地更新邻域区域的大小。聚类过程的目标函数通过欧几里得距离和自适应计算的 Hausdorff 距离的组合进行更新。通过与三位专家的手动分割进行比较来评估分割结果。还将结果与基于核诱导距离的具有空间约束的 FCM 方法、无自适应区域选择的方法和常规 FCM 方法进行了比较。
对 30 例患者图像进行分割的结果表明，自适应方法的灵敏度、特异性、Jaccard 相似性和 Dice 系数分别为 93.60±5.33%、97.83±2.17%、86.38±5.80%和 92.58±3.68%。基于区域的平均对称表面距离（ASSD）、均方根对称距离（RMSD）和最大对称表面距离（MSSD）的区域度量分别为 0.03±0.04mm、0.04±0.03mm 和 1.18±1.01mm。除灵敏度外，所有指标均优于非自适应算法和常规 FCM。只有三个基于区域的度量指标优于基于核诱导距离的具有空间约束的 FCM。
在分割 US 图像时自适应地包含像素邻域信息可提高分割性能。结果表明，该方法在乳腺肿瘤 CAD 和其他 US 引导程序中具有潜在的应用前景。
